feat(v5-g05): sharding axis_assignments → extras.sharding_applied

Closes V5-G05 / cppmega-mlx-8be. UI ShardingTab selection was
propagation-decorative for stage_train. v5 now reports
extras.sharding_applied = {axis_assignments[], shard_dim,
microbatch_size, compile_mode} so e2e can prove UI selection
reaches train context. Actual distributed math is single-device for
stage_train; real multi-device train deferred to v6.

1/1 e2e green; 52 stage_train pytest unchanged.

Author: apstenku123

cppmega_v4/runner/stages.py

@@ -707,10 +707,33 @@ def _count(tree: Any) -> int:
         fp8_active = False
         # Wire fp8_enabled from spec.sharding (payload pydantic model)
         ws_sharding = getattr(ctx.spec, "sharding", None)
+        sharding_applied: dict[str, Any] | None = None
         if ws_sharding is not None:
             fp8_active = bool(getattr(ws_sharding, "fp8_enabled", False))
             if fp8_active:
                 train_dtype = "fp8"
+            # G05: surface axis_assignments so e2e can prove UI sharding
+            # selection reaches stage_train (math is single-device for
+            # stage_train; real distributed train is v6).
+            axes = getattr(ws_sharding, "axis_assignments", None) or []
+            if axes:
+                shard_dim = 1
+                axis_list = []
+                for a in axes:
+                    deg = int(getattr(a, "degree", 1))
+                    shard_dim *= deg
+                    axis_list.append({
+                        "axis_name": str(getattr(a, "axis_name", "")),
+                        "kind": str(getattr(a, "kind", "")),
+                        "degree": deg,
+                    })
+                sharding_applied = {
+                    "axis_assignments": axis_list,
+                    "shard_dim": shard_dim,
+                    "microbatch_size": max(1, batch // shard_dim),
+                    "compile_mode": str(getattr(
+                        ws_sharding, "compile_mode", "off")),
+                }
 
         # G06: memory peak instrumentation — bracket train loop with
         # reset_peak_memory + get_peak_memory; extras.memory_peak_bytes.
@@ -894,6 +917,7 @@ def _count(tree: Any) -> int:
                 "graph_diff": graph_diff,
                 "gradient_clip": clip_extras,
                 "memory_peak_bytes": memory_peak_bytes,
+                "sharding_applied": sharding_applied,
                 "train_dtype": train_dtype,
                 "master_dtype": master_dtype,
                 "fp8_active": fp8_active,

vbgui/e2e/scenarios/34_sharding_apply.spec.ts

@@ -0,0 +1,31 @@
+// G05: sharding axis_assignments propagate to extras.sharding_applied.
+
+import { test, expect } from "@playwright/test";
+import { gotoApp, selectPreset, closeModal } from "../fixtures";
+
+test("G05: INITIAL_SPEC fsdp2 axis surfaces in extras.sharding_applied",
+  async ({ page }) => {
+    test.setTimeout(60_000);
+    await gotoApp(page);
+    await selectPreset(page, "llama3_8b");
+    // Topology must match axis degree
+    await page.getByTestId("topology-selector").selectOption("h100_8x");
+    await page.waitForTimeout(500);
+    await page.getByTestId("run-pipeline-toggle").click();
+    await page.getByTestId("run-pipeline-train").click();
+    const modal = page.getByTestId("run-result-modal");
+    await modal.waitFor({ timeout: 60_000 });
+    await page.getByTestId("run-result-expand-train").click();
+
+    const shardDim = parseInt(
+      (await page.getByTestId(
+        "run-result-extras-train-sharding_applied-shard_dim")
+        .textContent()) ?? "0", 10);
+    const compile = await page.getByTestId(
+      "run-result-extras-train-sharding_applied-compile_mode")
+      .textContent();
+    // INITIAL_SPEC has fsdp2 degree=8
+    expect(shardDim).toBe(8);
+    expect(compile?.trim()).toMatch(/regional|off|whole_model/);
+    await closeModal(page);
+  });